ai: collapse llamacpp into LLM_BACKEND env switch

Reverts the per-request backend="llamacpp" value. Chat/vision/embedding
backend is now a deploy-time decision (LLM_BACKEND=ollama|llamacpp),
applied globally across chat, vision describe, and embeddings — so
embedding vectors stay in one space across the index.

- Per-request backend whitelist back to "local"|"hybrid". A request
  arriving with backend="llamacpp" is rejected.
- LLM_BACKEND=llamacpp swaps the entire local stack to llama-swap:
  chat hits the chat slot, describe hits the vision slot, embeddings
  hit the embed slot. Hybrid mode still routes chat to OpenRouter
  but uses LLM_BACKEND for the describe pass.
- Drops env vars HYBRID_VISION_BACKEND, LLAMA_SWAP_VISION_MODELS,
  EMBEDDING_BACKEND (the last never shipped). Drops the
  LlamaCppClient.vision_models allowlist — capability inference now
  reports has_vision only for the configured vision_model slot.
- Drops the /insights/llamacpp/models handler. /insights/models is
  the single endpoint; returns Ollama servers under LLM_BACKEND=ollama
  and llama-swap slots (from LLAMA_SWAP_ALLOWED_MODELS) under
  LLM_BACKEND=llamacpp. Same envelope shape either way.
- New ai::embed_one helper routes embeddings through llama-swap when
  LLM_BACKEND=llamacpp (else Ollama). Wires it into the four
  insight_generator embedding sites.
- Cross-replay matrix simplifies to pre-llamacpp shape (local↔local,
  hybrid↔hybrid, hybrid→local allowed; local→hybrid rejected).

src/ai/llamacpp.rs

@@ -11,10 +11,10 @@
 //   `model` field, which is how llama-swap selects which backend process to
 //   run.
 // - `/v1/models` returns only the configured slot ids — capabilities aren't
-//   reported by the API, so `vision_models` is a config-time allowlist (env
-//   `LLAMA_SWAP_VISION_MODELS`) used to set `has_vision` on responses.
-//   `has_tool_calling` is assumed true for every slot, since llama-swap entries
-//   default to launching llama-server with `--jinja`.
+//   reported by the API, so we infer `has_vision` from a single config field
+//   (`vision_model`, defaulting to `"vision"`) and assume `has_tool_calling`
+//   is true for every slot, since llama-swap entries default to launching
+//   llama-server with `--jinja`.
 //
 // First consumer lands alongside the three-way backend dispatch in
 // insight_generator / insight_chat.
@@ -50,16 +50,10 @@ pub struct LlamaCppClient {
     /// Embedding model slot id (e.g. `"embed"`). Used for
     /// `generate_embeddings`.
     pub embedding_model: String,
-    /// Vision model slot id (e.g. `"vision"`). Used for `describe_image` and
-    /// included in `vision_models` automatically so capability lookups for
-    /// the default vision slot report `has_vision = true` even when the env
-    /// allowlist is empty.
+    /// Vision model slot id (e.g. `"vision"`). Used for `describe_image`,
+    /// and the only slot that reports `has_vision = true` in capability
+    /// lookups (llama-swap's `/v1/models` doesn't surface modality).
     pub vision_model: String,
-    /// Operator-curated set of slot ids known to be multimodal. Drives the
-    /// `has_vision` field in `list_models` / `model_capabilities`, since
-    /// llama-swap's `/v1/models` doesn't report modality. Empty allowlist
-    /// still marks `vision_model` as vision-capable.
-    pub vision_models: Vec<String>,
     num_ctx: Option<i32>,
     temperature: Option<f32>,
     top_p: Option<f32>,
@@ -83,7 +77,6 @@ impl LlamaCppClient {
             primary_model: primary_model.unwrap_or_else(|| DEFAULT_PRIMARY_MODEL.to_string()),
             embedding_model: DEFAULT_EMBEDDING_MODEL.to_string(),
             vision_model: DEFAULT_VISION_MODEL.to_string(),
-            vision_models: Vec::new(),
             num_ctx: None,
             temperature: None,
             top_p: None,
@@ -100,10 +93,6 @@ impl LlamaCppClient {
         self.vision_model = model;
     }
 
-    pub fn set_vision_models(&mut self, models: Vec<String>) {
-        self.vision_models = models;
-    }
-
     pub fn set_num_ctx(&mut self, num_ctx: Option<i32>) {
         self.num_ctx = num_ctx;
     }
@@ -692,7 +681,7 @@ impl LlamaCppClient {
             .and_then(|v| v.as_str())
             .unwrap_or_default()
             .to_string();
-        let has_vision = name == self.vision_model || self.vision_models.iter().any(|v| v == &name);
+        let has_vision = name == self.vision_model;
         // Tool calling is the default for llama-swap entries we configure
         // (--jinja flag); no negative-list mechanism yet, so report true.
         ModelCapabilities {
@@ -954,25 +943,21 @@ mod tests {
     }
 
     #[test]
-    fn capability_inference_uses_vision_model_and_allowlist() {
+    fn capability_inference_marks_only_vision_slot() {
         let mut c = LlamaCppClient::new(None, Some("chat".into()));
         c.set_vision_model("vision".into());
-        c.set_vision_models(vec!["qwen-vl".into()]);
 
         let m_chat = json!({ "id": "chat" });
         let m_vision = json!({ "id": "vision" });
-        let m_qwen = json!({ "id": "qwen-vl" });
         let m_other = json!({ "id": "embed" });
 
         let chat = c.parse_model_capabilities(&m_chat);
         let vision = c.parse_model_capabilities(&m_vision);
-        let qwen = c.parse_model_capabilities(&m_qwen);
         let other = c.parse_model_capabilities(&m_other);
 
         assert!(!chat.has_vision);
         assert!(chat.has_tool_calling);
         assert!(vision.has_vision);
-        assert!(qwen.has_vision);
         assert!(!other.has_vision);
     }
 }